A method for constructing, training, and applying a price spread direction prediction model

Abstract

The present invention discloses a method for constructing, training, and applying a spread direction prediction model. A multi-algorithm model under different scenarios is established, and finally, through an automated evaluation and screening mechanism between models, the best algorithm model is provided for spread direction prediction. In order to provide a more reliable spread reference basis for spot declaration, the present invention does not adopt a prediction scheme for the size of the spread. Instead, on the premise of compatibly considering the size of the spread, a more effective spread direction prediction scheme is adopted, and in combination with the loss function model constructed by the present invention, the reliability and accuracy of spread direction prediction are greatly improved.

Description

Technical Field

[0001] The present invention belongs to the technical field of power market data information processing, and particularly relates to a method for constructing, training, and applying a price difference direction prediction model. Background Art

[0002] In the power spot market, two market players, namely new energy power generation stations and electricity sales companies, will pay more attention to the price difference between the day-ahead clearing price and the real-time clearing price. For new energy power generation stations, the station traders need to declare the electricity quantity according to the predicted power generation situation of the station in the short term in the future, which is used as the standard for assessment and calculation of settlement income; similarly, for electricity sales companies, they also need to declare the short-term electricity consumption situation according to the electricity consumption prediction of their own users, and this declaration is also used as the standard for assessment and calculation of settlement income. Since the accuracy of grasping the size of the price difference between the day-ahead clearing price and the real-time clearing price in the short term in the future directly affects the level of the final assessment and settlement income, being able to accurately predict the size and direction of the price difference will naturally become the most crucial factor.

[0003] The existing technologies mainly use the disclosed data provided by the spot market, and artificially specify data features according to the trading experience of the traders themselves. Based on these data features, an algorithm model that can predict the price difference between the day-ahead clearing price and the real-time clearing price is trained through a pre-established fixed algorithm model, and this model is used to predict the price difference in the short term in the future. Finally, the spot electricity quantity declaration is made based on the predicted result of this price difference.

[0004] The disadvantages of the existing technologies mainly exist in the following three aspects. First: In extracting effective feature data from a large amount of disclosed data, clearing data, and meteorological data, the existing technologies still rely on the experience of the traders themselves, so the long-term effectiveness and correctness of the features cannot be guaranteed, and the artificially specified features cannot be guaranteed to be applicable to different seasons, different holidays, different windy seasons, and different weather conditions. Second: Most of the algorithm models of the existing technologies rely on a pre-constructed algorithm model, so the generalization ability and universality of this model cannot be guaranteed. The most important manifestation is that a single model can only ensure its effectiveness under a certain climate condition (such as the climate conditions in summer and autumn) and a certain variation rule (such as the overall network load is relatively stable and the wind and light output is relatively gentle). Third: The existing technologies mainly focus on the prediction of the size of the price difference, but in most cases, the size of the price difference changes erratically, the price difference amplitude is unstable, and the time period for the formation of the price difference is unpredictable. As a result, the prediction accuracy of the size of the price difference is very low, and thus it cannot guarantee to provide a reliable, stable, and effective reference basis for the trading entities in the spot market.

[0005] Therefore, we propose a method for constructing, training, and applying a price difference direction prediction model to solve the above problems. Summary of the Invention

[0006] The technical problem to be solved by the present invention is to provide a method for constructing, training, and applying a price spread direction prediction model, aiming to fully realize the automated data feature extraction functional module, so as to ensure that the extracted data features can meet different seasons, different holidays, different large and small wind seasons, and different weather conditions; in order to further improve the generalization ability and universality of the algorithm model, the present invention will establish multi-algorithm models under different scenarios, and finally provide the best algorithm model for price spread direction prediction through the automated evaluation and screening mechanism between models; in order to provide a more reliable price spread reference basis for spot declaration, the present invention does not adopt the prediction scheme of price spread size, but adopts a more effective price spread direction prediction scheme on the premise of considering the price spread size, and combines the loss function model constructed by the present invention to greatly improve the reliability and accuracy of price spread direction prediction.

[0007] The object of the present invention is achieved by the following technical solutions:

[0008] A price spread direction prediction model, which includes a feature screening algorithm model module, a prediction algorithm model module, and a model screening algorithm model module,

[0009] The feature screening algorithm model module includes an adaptive feature screening algorithm model;

[0010] The prediction algorithm model module includes a multi-scenario neural network algorithm model and a high-stability loss function algorithm model;

[0011] The model screening algorithm model module includes a strong intelligent model screening algorithm model and.

[0012] Another object of the present invention is to provide a method for training a price spread direction prediction model. The feature screening algorithm model module obtains disclosed data, clearing data through the spot trading center platform, and meteorological data through the European Meteorological Center, and then uses the adaptive feature screening algorithm model to extract the best feature data; the prediction algorithm model module is based on the screened data features, and forms the core of the prediction algorithm model by establishing a multi-scenario neural network algorithm model that meets different scenarios and combining a custom and high-stability loss function algorithm model applicable to the spot declaration scenario; the model screening algorithm model module mainly screens out the best model from multiple trained models through an intelligent model screening algorithm model.

[0013] Further: The above training method specifically includes the following steps:

[0014] S1. In the feature screening algorithm model module, appropriate feature data is obtained from numerous disclosure data, clearance data, and meteorological data as the input features of the neural network model. Meanwhile, the combination of the feature fusion, feature dimension elevation sub-algorithm model and the Kendall correlation coefficient analysis sub-algorithm model can screen out more relevant adaptive data features.

[0015] S2. In the prediction algorithm model module, based on the data features screened in step S1, by establishing and using a multi-scenario neural network algorithm model, and then training in combination with a loss function that comprehensively considers the size and direction of price spreads, a high-stability loss function algorithm model can be obtained.

[0016] S3. In the model screening algorithm model module, since multiple neural algorithm network models applicable to different scenarios are designed in the second step, a corresponding number of training models will also be obtained through training and learning on the screened data features. Using the recent feature data and clearance data as the test set, all models are cyclically used to predict this test set, and the corresponding price spread direction accuracy is obtained using the designed price spread direction prediction accuracy algorithm. Finally, the algorithm model with the highest accuracy is selected and used as the best model for predicting the future short-term price difference direction.

[0017] Further, the feature fusion and feature dimension elevation sub-algorithm model is specifically as follows:

[0018] d t =n t -w t -p t -l t

[0019] Where:

[0020] d t represents the non-new energy power generation space at the t-th moment;

[0021] n t represents the total network load demand at the t-th moment;

[0022] w t represents the total network wind power output at the t-th moment;

[0023] p t represents the total network photovoltaic power output at the t-th moment;

[0024] l t represents the transmission size of the tie line at the t-th moment.

[0025] Further, the Kendall correlation coefficient analysis sub-algorithm model is specifically as follows:

[0026]

[0027]

[0028] Wherein:

[0029] τ represents the magnitude of the Kendall correlation coefficient, and its value range is [-1, 1];

[0030] τ = -1 indicates that the two vectors are completely negatively correlated;

[0031] τ = 0 indicates that the two vectors have no correlation;

[0032] τ = 1 indicates that the two vectors are completely positively correlated;

[0033] n represents the sample size;

[0034] n a represents the number of concordant pairs in the comparison of two variables;

[0035] n b represents the number of discordant pairs in the comparison of two variables;

[0036] t i represents the number of clusters of the corresponding comparison vector when the price difference is at the i-th clustering level; u j represents the number of clusters of the corresponding price difference when the comparison variable is at the j-th clustering level. Further, the high-stability loss function algorithm model is specifically as follows:

[0037]

[0038] loss = diff_loss + log_loss

[0039] Wherein:

[0040] diff_loss represents the price difference magnitude and direction loss function;

[0041] log_loss represents the price difference direction logarithmic loss function;

[0042] represents the predicted value of the price difference direction at the t-th moment;

[0043] y t represents the true value of the price difference direction at the t-th moment;

[0044] p riqian,t represents the day-ahead clearing electricity price at the t-th moment;

[0045] p shishi,t represents the real-time clearing electricity price at the t-th moment;

[0046] It means that the day-ahead predicted electricity price at the t-th moment is greater than the real-time predicted electricity price;

[0047] It means that the day-ahead predicted electricity price at the t-th moment is equal to the real-time predicted electricity price;

[0048] It means that the day-ahead predicted electricity price at the t-th moment is less than the real-time predicted electricity price.

[0049] Furthermore, the formula of the algorithm model with the highest accuracy is as follows:

[0050]

[0051] Where:

[0052] It represents the predicted value of the spread direction at the t-th moment;

[0053] It means that the day-ahead clearing electricity price at the t-th moment is greater than the day-ahead real-time electricity price;

[0054] It means that the day-ahead clearing electricity price at the t-th moment is equal to the day-ahead real-time electricity price;

[0055] It means that the day-ahead clearing electricity price at the t-th moment is less than the day-ahead real-time electricity price;

[0056] p riqian,t It represents the day-ahead clearing electricity price at the t-th moment;

[0057] p shishi,t It represents the real-time clearing electricity price at the t-th moment.

[0058] The third object of the present invention is to provide a method for training a spread direction prediction model, which is applied to predicting the spread direction in the short term in the future.

[0059] Compared with the prior art, the beneficial effects of the present invention are as follows:

[0060] The present invention mainly predicts the direction of the price difference between the day-ahead clearing price and the real-time clearing price, and compatibly considers the influence of the size of the price difference in the design of the algorithm model. This is fundamentally different from the way of directly predicting the size of the price difference and is more effective. In the spot market, the starting points of each participating entity are somewhat different, and the trading varieties are also not the same. However, each entity will focus on the direction of the price difference between the day-ahead clearing price and the real-time clearing price, especially during the periods with a larger price difference amplitude, because the assessment and settlement income generated during these periods account for a larger proportion. Combining the loss function algorithm logic and model constructed by the present invention can greatly improve the prediction accuracy during the periods with a larger price difference amplitude, which will largely solve the problem that the existing technology has a low prediction accuracy for the size of the price difference, especially the inaccurate prediction or even reverse prediction during the periods with a larger price difference amplitude.

[0061] In order to ensure that the price difference direction prediction algorithm model has stable, reliable, and accurate prediction effects in different seasons, different holidays, different large and small wind seasons, and different weather conditions, the present invention constructs different algorithm network models, and designs a reasonable model evaluation mechanism and screening logic to realize the independent evaluation and screening of the algorithm model in different scenarios, so as to ensure the freedom, stability, and efficiency of the model.

[0062] At the same time, in order to improve the selection of data characteristics that are more in line with the actual situation from the disclosure data and clearing data published in the spot market, the present invention realizes the automatic and intelligent selection of data characteristics through algorithms such as feature fusion, feature dimension elevation, and correlation analysis. Thus, it avoids the problems of limitations, instability, and lack of universality of the selected features caused by the lack of artificial experience. BRIEF DESCRIPTION OF THE DRAWINGS

[0063] Figure 1 is the overall flowchart of the present invention;

[0064] Figure 2 is the flowchart of the feature screening algorithm model in the present invention;

[0065] Figure 3 is the flowchart of the prediction algorithm model in the present invention;

[0066] Figure 4 is the schematic diagram of the multi-scenario algorithm model in the present invention;

[0067] Figure 5 is the flowchart of the model screening algorithm model in the present invention. DETAILED DESCRIPTION OF THE EMBODIMENTS

[0068] The following will detail the implementation manners of the present application in conjunction with the drawings and embodiments, so as to fully understand the implementation process of how the present application uses technical means to solve technical problems and achieve technical effects and implement accordingly.

[0069] The present invention proposes an adaptive, multi-scenario, highly stable, and highly intelligent algorithm logic model for predicting the price difference direction between the day-ahead clearing price and the real-time clearing price. The overall design idea and process of this algorithm model are as follows Figure 1 shown. Among them, the feature screening algorithm model module obtains the disclosed data, clearing data through the spot trading center platform, and obtains meteorological data through the European Centre for Medium-Range Weather Forecasts (ECMWF). Then, it uses an adaptive feature screening algorithm model to extract the best feature data; the prediction algorithm model module is based on the screened data features. By establishing multiple neural network algorithm models that meet different scenarios and combining a custom, highly stable loss function algorithm model applicable to the spot declaration scenario, the core of the prediction algorithm model is jointly formed; the model screening algorithm model module mainly selects the best model from multiple trained models through an intelligent model screening algorithm model, and finally applies this model to the prediction of the price difference direction in the short term in the future.

[0070] Step 1: In the feature screening algorithm model module, it is necessary to obtain appropriate feature data from numerous disclosed data, clearing data, and meteorological data as the input features of the neural network model. Reasonable and effective feature selection directly affects the learning effect of the model and the prediction result of the final model. In the spot market trading center, the disclosed data that can generally be obtained includes data such as the whole network load forecast data, the whole network wind power output forecast data, the whole network photovoltaic output forecast data, and the tie line forecast data; the clearing data that can be obtained includes the day-ahead clearing price data and the real-time clearing price data. At the same time, relatively accurate meteorological data can be obtained through the European Centre for Medium-Range Weather Forecasts (ECMWF). However, considering that the data types are still relatively few, the present invention proposes a sub-algorithm model for feature fusion and feature dimension elevation. And in order to extract more effective and accurate data features, and considering the situation of data mutations (suddenly high or low) and outliers in the price difference data, the present invention adopts a sub-algorithm model for Kendall correlation coefficient analysis. Finally, through the combination of the two algorithm sub-modules, more relevant adaptive data features can be screened out. The detailed design idea and process of this algorithm module are as follows Figure 2 shown.

[0071] (1) Feature fusion and feature dimension elevation sub-algorithm model

[0072] According to the laws of the spot market and the supply-demand relationship that generally exists in nature, it is easy to obtain that when the power generation is greater than the power consumption, it will surely lead to a decrease in the electricity price; similarly, when the power generation is less than the power consumption, it will surely lead to an increase in the electricity price. Considering this obvious natural law, the present invention designs a feature fusion sub-algorithm model to reflect the reasonable degree of supply-demand relationship and uses this data as one of the data features. The specific algorithm model is as follows

[0073] d t = n t - w t - p t - l t

[0074] where:

[0075] d t represents the non - new - energy power generation space at the t - th moment;

[0076] n t represents the total network load demand at the t - th moment;

[0077] w t represents the total network wind power output at the t - th moment;

[0078] p t represents the total network photovoltaic power output at the t - th moment;

[0079] l t represents the transmission size of the tie line at the t - th moment;

[0080] (2) Kendall Correlation Coefficient Analysis Sub - algorithm Model

[0081] Since there must be a difference between the predicted and disclosed data and the actual operation data, there is generally a certain price difference between the day - ahead clearing price and the real - time clearing price. However, due to many situations such as sudden weather changes, sudden changes in load demand, and abnormal unit output, problems such as abnormal fluctuations and frequent alternations of the price difference will occur. Because the abnormal difference can also reflect the market rules and situations, in order to better utilize these abnormal data, the present invention first performs 10 - class clustering operations on the price difference data at each moment, the disclosed data of each variable, and the meteorological data at the corresponding moment through the k - nearest neighbor algorithm, and then obtains the corresponding clustering levels. Finally, the Kendall correlation coefficient analysis sub - algorithm model is used to obtain the characteristic variable data with a correlation greater than 0.6. The specific algorithm model is as follows

[0082]

[0083] where:

[0084] τ represents the magnitude of the Kendall correlation coefficient, and its value range is [-1, 1];

[0085] τ = - 1 indicates that the two vectors are completely negatively correlated;

[0086] τ = 0 indicates that the two vectors have no correlation;

[0087] τ = 1 indicates that the two vectors are completely positively correlated;

[0088] n represents the sample size;

[0089] n a represents the number of concordant pairs in the comparison of two variables;

[0090] n b represents the number of discordant pairs in the comparison of two variables;

[0091] t i represents the number of clusters of the corresponding comparison vector when the price difference is at the i-th clustering level;

[0092] u j represents the number of clusters of the corresponding price difference when the comparison variable is at the j-th clustering level;

[0093] Step 2: In the prediction algorithm model module, based on the data features screened in the first step, and by establishing neural algorithm network models applicable to different scenarios, such as algorithm models applicable to different wind seasons, algorithm models applicable to different levels of load demand, and algorithm models applicable to different weather conditions, and then combining the loss function designed in this patent that takes into account both the magnitude and direction of the price difference for training, training models for different scenarios can be obtained. The detailed design idea and process of this algorithm module are as Figure 3 shown.

[0094] (1) Multi-scenario algorithm model

[0095] In order to adapt to neural network algorithm models for different scenarios such as different wind seasons, different levels of load demand, and different weather conditions, the present invention proposes to design neural network algorithm models that can each exert their own network advantages in different scenarios and different dimensions. The specific implementation mechanism and algorithm principle of this algorithm model are as Figure 4 shown.

[0096] (2) High-stability loss function algorithm model

[0097] A reasonable, efficient, universal, and stable loss function is the key to the excellent learning effect of the algorithm model. In the spot market scenario, each participating entity pays more attention to the price difference between the day-ahead clearing price and the real-time clearing price. However, the existing technologies have poor prediction effects and low accuracy for the magnitude of the price difference. The present invention proposes a loss function that takes into account both the magnitude and direction of the price difference. By designing a reasonable compatibility method, two different loss functions are fused to obtain a high-stability loss function model. The specific implementation mechanism and algorithm principle of this algorithm model are as follows

[0098]

[0099] loss = diff_loss + log_loss

[0100] where:

[0101] diff_loss represents the loss function of the price difference size and direction;

[0102] log_loss represents the logarithmic loss function of the price difference direction;

[0103] represents the predicted value of the price difference direction at the t-th moment;

[0104] y t represents the true value of the price difference direction at the t-th moment;

[0105] p riqian,t represents the day-ahead clearing electricity price at the t-th moment;

[0106] p shishi,t represents the real-time clearing electricity price at the t-th moment;

[0107] represents that the day-ahead predicted electricity price at the t-th moment is greater than the real-time predicted electricity price;

[0108] represents that the day-ahead predicted electricity price at the t-th moment is equal to the real-time predicted electricity price;

[0109] represents that the day-ahead predicted electricity price at the t-th moment is less than the real-time predicted electricity price;

[0110] Step 3: In the model screening algorithm model module, since a variety of neural algorithm network models applicable to different scenarios are designed in the second step, the corresponding number of training models will also be obtained through training and learning on the screened data features. In order to select the training model with the best prediction effect for future short-term prediction from these models, the present invention designs a model screening algorithm that conforms to the spot declaration scenario. The specific algorithm model is as follows

[0111]

[0112] where:

[0113] represents the predicted value of the price difference direction at the t-th moment;

[0114] represents that the day-ahead clearing electricity price at the t-th moment is greater than the day-ahead real-time electricity price;

[0115] represents that the day-ahead clearing electricity price at the t-th moment is equal to the day-ahead real-time electricity price;

[0116] represents that the day-ahead clearing electricity price at the t-th moment is less than the day-ahead real-time electricity price;

[0117] p riqian,t represents the day-ahead clearing electricity price at the t-th moment;

[0118] p shishi,t represents the real-time clearing electricity price at the t-th moment;

[0119] The specific implementation principle of this algorithm module is as follows: taking the feature data and clearing data of recent days as the test set, using all models to predict this test set in a loop, obtaining the corresponding accuracy rate of the price difference direction by using the designed accuracy rate algorithm for predicting the price difference direction, finally selecting the algorithm model with the highest accuracy rate, and using this model as the best model for predicting the future short-term electricity price difference direction. The detailed design idea and process of this algorithm module are as Figure 5 shown.

[0120] In summary, it is only a preferred embodiment of the present invention, and does not impose any form of limitation on the present invention. Any simple modification and equivalent change made to the above embodiments based on the technical essence of the present invention shall fall within the protection scope of the present invention.

[0121] The above description is only a preferred embodiment of the present invention, and does not impose any form of limitation on the present invention. Any simple modification, equivalent change and modification made to the above embodiments based on the technical essence of the present invention shall fall within the scope of the technical solution of the present invention.

Claims

1. A method for training a price spread direction prediction model, characterized in that: The model includes a feature screening algorithm model module, a prediction algorithm model module, and a model screening algorithm model module. The feature screening algorithm model module includes an adaptive feature screening algorithm model. The prediction algorithm model module includes a multi-scenario neural network algorithm model and a high-stability loss function algorithm model. The high-stability loss function algorithm model includes a multi-scenario neural network algorithm model, and the loss function of the price difference size and direction is trained using the multi-scenario neural network algorithm model. The model screening algorithm model module includes a strong intelligent model screening algorithm model. Specifically, it includes the following steps: S1. In the feature screening algorithm model module, appropriate feature data is obtained from a large number of disclosure data, clearing data, and meteorological data as the input features of the neural network model. At the same time, the combination of the feature fusion, feature dimension elevation sub-algorithm model and the Kendall correlation coefficient analysis sub-algorithm model can screen out more relevant adaptive data features. S2. In the prediction algorithm model module, based on the data features screened in step S1, a multi-scenario neural network algorithm model is established and then trained in combination with a loss function that comprehensively considers the price difference size and direction, and a high-stability loss function algorithm model can be obtained. S3. In the model screening algorithm model module, since multiple neural algorithm network models applicable to different scenarios are designed in the second step, a corresponding number of trained models will also be obtained through training and learning of the screened data features. Using the recent feature data and clearing data as the test set, all models are cyclically used to predict this test set, and the corresponding price difference direction accuracy is obtained using the designed price difference direction prediction accuracy algorithm. Finally, the algorithm model with the highest accuracy is selected and used as the best model for predicting the future short-term price difference direction. The specific content of the feature fusion and feature dimension elevation sub-algorithm model is as follows: d t = n t - w t - p t - l t Where: d t represents the non-new energy power generation space at the t-th moment; n t represents the network-wide load demand at the t-th moment; w t represents the total wind power output of the whole network at the t-th moment; p t represents the total PV output of the entire network at the t-th moment; l t Indicates the transmission magnitude of the tie line at the t-th moment.

2. The method for training the spread direction prediction model according to claim 1, wherein: The specific content of the Kendall correlation coefficient analysis sub-algorithm model is as follows: Where: τ represents the magnitude of the Kendall correlation coefficient, and its value range is [-1, 1]. τ = -1 indicates that the two vectors are completely negatively correlated. τ = 0 indicates that the two vectors have no correlation. τ = 1 indicates that the two vectors are completely positively correlated. n represents the sample size. n a represents the number of concordant pairs in the comparison of two variables; n b represents the number of dissonant pairs in the comparison of two variables; t i represents the number of clusters of the corresponding comparison vectors when the price difference is at the i-th clustering level; u j Indicates the number of clusters corresponding to the price difference when the comparison variable is at the j-th clustering level.

3. The method for training the price difference direction prediction model according to claim 1, wherein: The specific content of the high-stability loss function algorithm model is as follows: loss = diff_loss + log_loss Where: diff_loss represents the price difference size and direction loss function. log_loss represents the price difference direction logarithmic loss function. represents the predicted value of the price difference direction at the t-th moment; y t represents the true value of the price difference direction at the t-th moment; p riqian,t represents the day-ahead clearing price at the t-th moment; p shishi,t represents the real-time clearing price at the t-th moment; Indicates that the predicted electricity price before the t-th moment is greater than the real-time predicted electricity price; It means that the predicted electricity price before the t-th moment is less than the real-time predicted electricity price.

4. The method for training the spread direction prediction model according to claim 1, wherein: The formula for the algorithm model with the highest accuracy is as follows: Where: represents the predicted value of the price difference direction at the t-th moment; Indicates that the day-ahead clearing price at the t-th moment is greater than the day-ahead real-time price; It means that the day-ahead clearing price at the t-th moment is equal to the day-ahead real-time price; It means that the day-ahead clearing price at the t-th moment is less than the day-ahead real-time price; p riqian,t represents the day-ahead clearing price at the t-th moment; p shishi,t represents the real-time clearing price at the t-th moment.

5. A method for training a price difference direction prediction model according to any one of claims 1-4 is applied to predict the price difference direction in the short term in the future.

Images